INST0098 Chap.2 Research and Writing with AI
Research and Writing with AI
Research and Writing with AI treats generated text as material that requires verification and judgement. The chapter separates useful assistance from unsupported authority, covering prompt design, citation checking, framing analysis, revision and disclosure.
It emphasises that a detailed prompt cannot compensate for missing evidence and that repeated model output does not validate a category or interpretation.
Divide AI-supported work into tasks with different risk. Brainstorming search vocabulary is not the same as summarising an unread article, and reorganising notes is not the same as making a factual claim.
Before prompting, state what the output may influence and what evidence will control the final decision. This boundary prevents a convenient draft from quietly becoming the authority for the project.
Use a claim ledger during verification. Give every factual or interpretive statement a status: located, supported, qualified, contradicted or unresolved. Record the source and the passage that justifies the status.
A real citation can still be misrepresented, so checking the title is insufficient. Read enough context to understand the population, method and qualification. For causal language, separate evidence that two events occurred from evidence that one produced the other, and search for an alternative explanation.
Framing requires a second audit.
List the categories the output assumes, the perspective treated as normal and the groups or possibilities that disappear. Compare those choices with the language of the corpus and concepts in the literature. Repeating a prompt with small changes can reveal instability, but stable repetition can also reproduce the same embedded assumptions.
Keep boundary cases and rejected classifications because they show where the scheme fails.
Add a red-team pass before revision. Ask what evidence would contradict the output, what source type would be more authoritative and which named entity, quotation or numerical claim is easiest to check independently. Search for the strongest counterexample rather than another supporting summary.
If the result changes when a prompt is reframed, record the unstable part and keep that uncertainty visible. This practice turns verification into an active attempt to refute a convenient answer instead of a search for reassurance.
When revising, rebuild the paragraph from checked notes instead of cosmetically editing generated prose. Decide the order of claims, evidence and qualification yourself.
Preserve a process record containing the tool, purpose, prompt, useful output, corrections and rejected material, subject to the assessment rules and privacy obligations. That record supports accurate disclosure and makes it possible to explain the intellectual work performed after generation.
End with a removal test. Delete the generated text and ask whether the evidence trail still supports every sentence in the draft.
If a conclusion cannot be reconstructed from sources and documented reasoning, it is not ready to remain. The goal is not to disguise assistance but to ensure that scholarly responsibility stays with the researcher.
Review the ledger after editing, because a polished transition can introduce a stronger claim than the verified note supports. Accuracy must survive the final prose, not only the research stage.
What this chapter covers
- 01
Bounded AI tasks
- 02
Prompt design
- 03
Citation verification
- 04
Framing audits
- 05
Research records
- 06
Transparent disclosure
Worked example · free
Audit a generated literature summary
- 2Locate and read each cited source rather than trusting the reference string.
- 2Separate evidence of policy change from evidence of behaviour.
- 2Record corrections, rejected claims and the tool's role in the process.
Key terms
- Hallucinated citation
- A reference generated in plausible form that does not accurately identify a real source.
- Verification trail
- A record connecting generated suggestions to checked sources, corrections and final use.
- Framing audit
- A review of categories, assumptions and omitted perspectives within an output.
Research and Writing with AI FAQ
Can a detailed prompt guarantee accuracy?
No. Specific instructions may improve relevance and format, but an output can remain false, unsupported or biased. Accuracy requires checking against reliable sources and the context in which those sources make their claims.
Why keep rejected AI output?
A record of rejected suggestions shows how the research developed and supports accurate disclosure. It also prevents a discarded error from returning later when notes are reorganised.
Does consistent output prove a category is valid?
Consistency can reflect stable model assumptions rather than the source material. Categories still need theoretical justification, corpus testing and attention to cases they exclude or distort.
Assessment move
Practise on a short scholarly passage. Ask a tool for search terms, a summary and a counterargument, then audit each result separately. Mark factual statements, interpretive statements and organisational suggestions in different colours. Verify citations in a library database, compare every summary with the passage and write down any missing qualification.
Rephrase the prompt to expose instability, but remember that repeated answers are not proof. Build a process note containing the task, prompt, output used, correction made and material rejected. Finally rewrite the useful material in an argument structure you control. This routine develops the habit of using assistance without transferring scholarly responsibility to the tool.
Working through Research and Writing with AI in INST0098? Sia is AskSia’s AI Arts and Humanities tutor — ask any INST0098 Research and Writing with AI question and get a clear, step-by-step explanation grounded in how INST0098 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.